SCALE is a GPGPU programming toolkit designed to compile CUDA applications for AMD GPUs without requiring modifications to the original CUDA program or its build system, aiming to broaden CUDA's hardware compatibility.
Codestral Mamba
Codestral Mamba, a new AI model by Mistral AI, offers linear time inference and can handle sequences of infinite length, diverging from traditional Transformer models to enhance code productivity. Read more about Mamba models
Guide to Machine Learning with Geometric, Topological, and Algebraic Structures
Modern machine learning is evolving to embrace non-Euclidean data, characterized by complex geometric, topological, and algebraic structures, diverging from the traditional Euclidean geometry foundation.
XLSTMTime: Long-Term Time Series Forecasting with xLSTM
xLSTMTime, an adaptation of the extended LSTM (xLSTM) architecture for long-term time series forecasting (LTSF), outperforms current models, including those based on transformers and LTSF-Linear, by incorporating exponential gating and a revised memory structure.
Transformer Layers as Painters
Empirical studies on frozen pretrained transformers reveal distinct behaviors between lower, middle, and final layers, with middle layers exhibiting uniformity.
\[R\] Protein language models expose viral mimicry and immune escape
Protein Language Models (PLMs) achieve a 99.7% accuracy in distinguishing viral proteins from human ones, highlighting their potential in identifying viral mimicry strategies and immune escape mechanisms.
\[R\] Discussion on the paper: Transcendence: Generative Models Can Outperform The Experts That Train Them
The paper, "Transcendence: Generative Models Can Outperform The Experts That Train Them," has sparked significant interest for its claim that generative models can surpass the expertise of their training datasets' contributors. Read the paper
\[R\] Any-Property-Conditional Molecule Generation with Self-Criticism using Spanning Trees
Spanning Tree-based Graph Generation (STGG) has been enhanced to STGG+, incorporating a Transformer architecture and random masking to generate molecules conditional on desired properties, significantly improving upon traditional SMILES and graph diffusion models.
🤗SmolLM - blazingly fast and remarkably powerful
SmolLM introduces a new benchmark in small language models with sizes of 135M, 360M, and 1.7B parameters, trained on a meticulously curated high-quality dataset, SmolLM-Corpus, which includes Cosmopedia v2, Python-Edu, and FineWeb-Edu, totaling over 252B tokens.
TwoMinutePapers - New AI: This Is A Gaming Revolution!
Embodied avatars in gaming now allow players to control their virtual characters with their own body movements, rather than traditional controllers, promising a more immersive experience.
\[D\] Best open source LLM for graph based questions answering
Open source Large Language Models (LLMs) are sought for two distinct tasks: creating knowledge graphs from unstructured text and answering questions based on these graphs.
\[N\] Found a helpful tool for scaling LLM on a large amount of unstructured data
Roe AI simplifies the process of scaling LLMs for extracting features from unstructured data such as PDFs and images/videos, bypassing the need for complex coding or inadequate platform support.
Foundational Autoraters: Taming Large Language Models for Better Automatic Evaluation
FLAMe, a Foundational Large Autorater Model, leverages a diverse collection of 5M+ human judgments across 100+ quality assessment tasks, outshining LLMs like GPT-4 and Claude-3 in generalization to various tasks. Read more
RUBICON: Evaluating conversations between humans and AI systems
RUBICON, an automated assessment technique introduced by Microsoft Research, transforms minimal datasets into extensive domain-specific rubrics, enhancing the evaluation of human-AI interactions in specialized settings like software development.
Data-driven model improves accuracy in predicting EV battery degradation
Microsoft Research and Nissan Motor Corporation have developed a machine learning method that predicts EV battery degradation with a remarkable average error rate of 0.94%, enhancing the accuracy of battery recycling efforts significantly.